How I built Drafted
A CV can leave out experience that matters.
I started building Drafted after seeing a chatbot mistake a gap in a CV for a gap in someone’s experience. The first version picked out what mattered in a job ad and asked the applicant about anything their CV left unclear.
Experience training new colleagues.
Prepared weekly stock reports.
Did you ever show a new colleague how to prepare those reports?
What if the next application could build on the last?
- Previous applicationsCVs, letters and answers
“I taught two new colleagues how to prepare the reports.”
- Career profileExperience from your answer
Trained two colleagues in weekly stock reporting.
- Next CV draftNew job ad asks for training experience
Trained two colleagues to prepare weekly stock reports.
Today, Drafted works with one CV at a time. I’m developing a career profile so it can reuse relevant experience and ask only about what’s missing.
How I built the research behind Drafted.
I wanted product decisions to draw on carefully reviewed research about job seekers and AI. I built the system around two permanent agents.
Research Discovery & Screening Agent
Found new sources, checked for duplicates and recorded the reason for each screening decision.
Handoff · Candidates selected for full-text reviewResearch Deep Review Agent
Read the papers, checked how studies were conducted and traced each finding to its source.
Handoff · Reviewed claims and their limitationsIRP Research Systems Architect
Defines what counts as evidence, how records are identified, where files belong and what automation may do.
Output · Workspace contractsManual or weekly
I can start a run myself. The scheduled task starts the same workflow each week. Both read the current research records.
Output · Start a research cycleIRP Research Orchestrator
Reads the workspace, identifies a gap and searches for evidence. Each run writes its results back and produces a short digest.
- Discover & screenLog each candidate and the screening decision.Advance → Full-text reviewHold or reject → Record the reasonCandidate register ·
research-candidates.md - Read & verifyCheck methods, sample, findings and limits. Promote verified sources to LIT or PRO records linked to the original candidate.Source ledger ·
research-ledger.md - InterpretLink proposed implications to specific passages, record conflicting evidence and assess how well the findings apply to Drafted.Workspace · Proposed implications and run digest
A run assigned interview IDs to hypothetical experiments. I used the Architect to correct the contract. An interview ID now requires a session that actually took place.
Source library
Permitted full-text PDFs for checking the original evidence.
Research history
Candidate records, reviewed sources, page references, proposed implications and run digests.
I decide what reaches Drafted.
I can turn a proposal into a design principle, a product decision or an experiment, or leave it open.
Proposed extension · Independent review
A separate reviewer would receive the proposed change, exact source passages and any conflicting evidence. It would return which claims the evidence supports, where the proposal goes too far and what to test next. I would review that response alongside the original proposal.
- Example proposal
- Remove the editor checkpoints to reduce friction.
- What the reviewer would challenge
- One tester wanted to see a result first. That gives me a reason to revisit when the first checkpoint happens. It leaves open how many checkpoints people need to feel in control.
- Experiment to consider
- Show a first draft before asking for approval. Compare how easily people can correct it and how much effort they report.
One finding made me reconsider the checkpoints I’d added to keep applicants in control. Each pause also cost them effort. During testing, one person told me they just wanted to see the result.